A Robust Fire Detection Model via Convolution Neural Networks for Intelligent Robot Vision Sensing.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Traditional fire detectors face environmental interference and limitations.
• Current deep learning fire detection methods struggle with dynamic fire changes and precision.

Purpose of the Study:

• To develop a more accurate and robust fire detection system using deep learning.
• To enhance fire identification capabilities in dynamic and varied environments.

Main Methods:

• Implemented K-mean++ for optimized anchor box clustering in YOLOv5.
• Integrated dynamic convolution into the YOLOv5 architecture.
• Performed pruning on YOLOv5's neck and head network layers to boost detection speed.

Main Results:

• The dynamic convolution YOLOv5 method significantly improved recall, precision, and F1-score compared to the standard YOLOv5.
• Achieved precision improvements of 13.7%, 10.8%, and 6.1% over three other deep learning methods.
• Demonstrated superior F1-scores, with improvements of 15.8%, 12%, and 3.8% respectively.

Conclusions:

• The proposed dynamic convolution YOLOv5 method offers superior fire detection performance.
• The system is effective for both short-range indoor and long-range outdoor fire identification.
• This approach addresses limitations of traditional detectors and existing deep learning models.